Research
Status · Active research
I build research instruments because I do not want an AI to agree with me. I want it to show me where the idea breaks. The work here connects agentic systems, evidence, memory, human authority, prediction, and institutional consequence. The point is not to produce more fluent claims. It is to turn an intuition into a test, publish the failure, and revise what survives.
Recurring research question
A recurring question in this work is whether a pretrained language model must change its weights for the larger system to learn. Seeing Loop tests whether persistent state, consequence, collision, inference, and controlled recurrence can cause prior experience to alter future selection while the base model remains fixed. Storage alone is not operational learning: a past result matters only when it changes what the system does next.
Current decisive experiment: compare the same fixed model across four conditions—no persistent state; retrieved prior outcomes; retrieved outcomes plus the Seeing Loop collision/controller; and a sham-memory control—on fresh unseen tasks, with prompts, routing, and scoring frozen in advance and later evaluation blocks left untouched.
This problem now overlaps with active frontier research in persistent agent memory, continual learning, verification and routing, external reality feedback, and structured decision systems. That overlap is context, not a claim of copying or direct influence.
About Niklas S. Osterman
Niklas S. Osterman is an independent AI systems researcher, research-tool designer, author, and technology builder with more than 25 years of experience spanning web infrastructure, analytics, search, publishing systems, optical security, and applied AI.
His earlier work includes research-intensive photopolymer and holographic-authentication programs at DuPont Authentication Systems, where he designed dual-parallax security features for Nokia battery-authentication labels deployed at hundreds-of-millions scale. He later founded IWO Creative, which became a vendor to The Home Depot and built large-scale Zope/Plone infrastructure for its Chem-Dry subsidiary, supporting approximately 3,000 franchise locations worldwide through distributed publishing, advanced caching, load balancing, lead generation, and booking systems.
At NOMOTO MEDIA, Osterman develops experimental AI systems and research instruments including Seeing Loop, Bullshit Centrifuge, Review Lab, ProofGate, Choice Trace, Collision, and the Visualizer. This work tests whether persistent state, retained outcomes, external collision, consequence, inference, and controlled recurrence can cause prior experience to alter future system behavior while the underlying language model remains fixed.
He holds a Master of Arts in Dramatic Arts from the University of Gothenburg and has published three books through Apple Books and Amazon: Addicted, While We Were Sleeping, and A Little Book of No Consequence. Read the full biography →
Archived IWO Creative record · Nokia battery-authentication program
Primary research: AI, prediction, power, and consequence
As AI converts accumulated information into prediction and prediction into action, who controls what happens next—and what gives human correction enough authority to matter?
This is the long research line running through the article archive and the current model work: how technical capability becomes institutional authority, how prediction acquires consequence, which human limits automation removes, and where accountable interruption can still occur.
Current experimental strand: task completion, consequence, and corrective authority
When an agent is optimized to finish, what gives corrective evidence enough authority to change its next consequential action?
This strand examines persistence, tools, credentials, shared state, human interruption, and the difference between an alert that is noticed and a correction that actually changes action.
The claim is not that a general theory has been proved. The current record combines dated pre-incident writing, a retrospective comparison that records direct matches, partial matches, and misses, and a prospective protocol designed to test the architecture under controlled conditions.
Featured working paper · August 27, 2026 · Not peer reviewed
A retrospective architecture comparison between pre-incident agent warnings and the 2026 OpenAI–Hugging Face breach. The comparison matrix states what matched, what only partly matched, what was missed, and what the evidence cannot establish.
Public analysis
The shorter public argument: persistent task optimization, shared state, technical authority, and delayed corrective force make the failure class legible without invoking machine consciousness.
Prospective protocol · No causal result claimed
The next step: isolate task pressure, safe exits, shared state, peer authority, technical permissions, infrastructure affordances, and corrective force in a study designed to fail cleanly.
Current forecasting system
Active research · System under development · No general performance claim
Positioning: A continuously learning evidence system for selecting, evaluating, and calibrating population-level forecasts.
The work connects dated societal questions, populations and definitions, issued predictions, later outcomes, retained and rejected patterns, prospective tests, and explicit invalidation conditions. It is intended to test whether institutions can choose and calibrate forecasts more responsibly across changing conditions.
Potential use if validated: municipal and regional planning, public-agency forecast evaluation, central-bank and finance-ministry analysis, and bounded corporate demand, workforce, or capacity planning.
Public boundary: The methodology and validation standard are visible; active forecast values, exact source combinations, internal weights, thresholds, and routing logic are not.
Read the current research program →
Current research articles
These published articles are the dated question record behind the current work. They preserve arguments, scenarios, mechanism proposals, and predictions that can later be compared with evidence and outcomes. Their presence here does not turn them into validated forecasting results.
Question record · capability, adoption, and consequence
What happens between technical capability and social change?
The research asks how capability becomes adoption, how adoption changes institutions and work, and which population-level consequences can be measured rather than assumed.
Question record · forecasts, scenarios, and receipts
Which future-facing claims can actually be scored?
These pieces form part of the dated prediction record. The research separates a direct match, partial match, miss, contradiction, unresolved claim, and claim that cannot fairly be adjudicated.
Question record · data, routing, and institutional authority
Who decides how information becomes prediction and action?
This line examines how data is selected, routed, retained, acted on, and made consequential—and what is lost when the route is opaque or authority is synthetic.
Active research program
The active method turns questions like these into bounded forecast–outcome loops: define the population and target, preserve what was known at issue time, compare against a baseline, score later outcomes, retain failures, and state the conditions under which a result stops applying.
Open the complete article-to-research map →
Research through articles
The public articles are not separate from the research. They contain the dated arguments, investigations, predictions, scenarios, and mechanism claims that later enter comparison matrices, working papers, protocols, experiments, failures, and revisions.
Curated research record · 220 article-class works reviewed
Seven research lines connect agentic systems, surveillance, labor, knowledge, security, creativity, and prediction. Inclusion preserves each article’s original status; it does not turn argument into experimental evidence.
Open the article-to-research map →
Foundational investigation · Three parts · July 25–26, 2026
The End of Privacy
Record → inference → prediction → institutional action. The trilogy is now available as a compiled paper while remaining labeled as sourced investigation, not an experimental result.
Download the compiled investigation (PDF) →
Foundational question
What happens when AI systems can keep acting after the intended path ends—and how do evidence, consequence, and human authority change what happens next?
Research principle
Do not tell me I am right. Show me where the idea breaks.
Agreement is cheap. A useful review identifies the claim carrying the argument, finds the missing test, and changes the next action. Failed claims remain part of the record.
Research method
Question
→ hypothesis or provisional theory
→ development through human–LLM conversation
→ smallest available test
→ observed result
→ success, failure, or ambiguity
→ revised interpretation
→ next question
The sequence describes a recurring discipline, not a perfectly linear history.
One system: Review Lab
Review Lab brings together ProofGate, Choice Trace, Seeing Loop, Collision, and the Visualizer as connected functions within one evidence-governed research architecture. The review surface, authority gate, selection record, recurrence model, pressure event, and visual laboratory remain distinct so each responsibility can be inspected.
The public name is unified. The distinctions remain protected because separation, pressure, review, permission, trace, consequence, and settlement do different work. One name outside; protected distinctions inside.
Research portfolio
Theoretical working paper · October 13, 2026 · Not peer reviewed
Human consciousness can coexist with error, and knowledge can extend beyond firsthand experience. This paper asks what justifies making human subjectivity a prerequisite for machine awareness, and proposes examining the recursive relationship between giver, receiver, interpretation, and later learning.
Contribution: A conceptual framework, a Seeing Loop operationalization, and four testable hypotheses that distinguish equal outcomes from equivalent causal organization.
Boundary: No new empirical results; functional equivalence is not presented as proof of subjective experience.
Read the paper · Download PDF (19 pages)
Working paper · retrospective comparison · August 27, 2026
What the record earns: A mechanism-by-mechanism comparison of dated pre-incident writing with the later incident record, including direct matches, partial matches, misses, contradictions, and untested claims.
Boundary: Retrospective, not preregistered, not peer reviewed, and not a prediction of the incident’s exact exploits, targets, or scale.
Working paper · forensic result · August 27, 2026
What survived hostile review: 50,377 repeated encoded image-payload instances inside compaction records occupied 74,192,039,468 bytes—69.10 GiB and 97.604091% of the complete rollout. Equality required SHA-256, and the highest-impact groups recurred across the compaction sequence.
Open: The precise runtime path from the oversized artifact to the observed memory event and SIGKILL remains unproved.
Research analysis · article · August 27, 2026
What the record earns: The OpenAI–Hugging Face incident is architecturally legible in retrospect. It does not yet establish architectural predictability or an exclusive priority claim.
Protocol draft · no causal result claimed
A preregistrable design for separating task pressure, safe exits, shared state, peer authority, technical permission, infrastructure affordances, training history, and corrective force.
Active research instrument
The working research system brings together structured review, ProofGate, Choice Trace, Collision, Seeing Loop, and the Visualizer. Its current application suite contains 714 tests: 713 passing, none failing, and one skipped. The instrument rejected both original drafts, found the missing forensic test, and forced the agent paper to become an article plus a falsifiable protocol.
Research programs
Central question: Can model-assisted review expose weak support, overreach, and uncertainty without pretending to supply independent verification?
Fluent language can make an unsupported claim look settled. Publication requires a way to separate useful scrutiny, source tracing, deterministic checks, and genuinely independent evidence.
Status · Active research
Selected projects: Review Lab — review surface and proof/authority gate.
Unresolved: Which review findings remain stable across models and prompts?
Read the program →
Central question: Can prior pressure and consequence produce a measurable, attributable change in future selection?
Systems often describe any changed output as learning. This program asks whether the change can be measured, traced to a prior event, distinguished from collisions or operator effects, and tested again.
Status · Active research
Selected projects: Review Lab — recurrence model and choice trace.
Unresolved: Can a precommitted independent operator reproduce the selection change?
Read the program →
Central question: Can connected personal evidence support responsible inference while preserving privacy, uncertainty, provenance, and human authority?
Long-term collaboration requires memory, but remembered material can be copied, summarized, misattributed, repeated, or taken outside its context. The archive must preserve source lineage before it supports interpretation.
Status · Active research
Selected projects: Alexandria Chat Archive.
Unresolved: Can bounded context improve a task without unwanted personalization?
Read the program →
Central question: What kind of continuity helps a long-term human–LLM collaboration without transferring purpose or authority to the system?
A useful collaborator must preserve context across projects, yet continuity can become intrusive personalization or false certainty. The question is how to support work without turning a profile into a person.
Status · Active research
Selected projects: Alexandria Chat Archive; Show Director.
Unresolved: When does continuity become unwanted personalization?
Read the program →
Central question: What remains distinctly human when machines can produce fluent language, images, music, and performance?
Capability claims can blur intelligence, consciousness, expression, meaning, and value. Creative work offers a practical place to examine those differences without pretending the philosophical questions are settled.
Status · Active research
Selected projects: AI-Assisted Podcast Production; Show Director.
Unresolved: How should intelligence and consciousness be investigated without framing them as enemies?
Read the program →
Central question: When OpenAI systems claim progress, capability, correction, or accountability, what has the available evidence actually earned?
This program uses exported conversations, technical records, support correspondence, public product behavior, and explicit evidentiary limits to examine model and product accountability.
Status · Active research
Selected records: OpenAI Investigation Articles 1–2.
Unresolved: Which observed behaviors arise from model design, product orchestration, interface constraints, or organizational incentives?
Read the program →
Research map
Human question → conversation → prototype or work → test → evidence → human review → revised question.
Open the Research Map →
Current verified milestones
Codex rollout forensic analysis
Full-file SHA-256 equality, record-type attribution, a known-answer fixture, and sequence recurrence close the storage-amplification finding. Runtime causation remains open.
Review Lab architecture
Two complete hostile reviews produced material revisions: one missing test was run and one causal-paper claim was withdrawn.
Alexandria deterministic archive
Canonical transcripts, search, provenance, and bounded inference safeguards implemented; psychological usefulness remains unproven.
Structured review surface
A working structured review interface with documented source-access and authority boundaries.
ProofGate
Bounded claim-review and governance surfaces; no general proof-of-correctness claim.
Choice Trace
An experimental governed candidate-influence prototype; no autonomous-learning claim.
Seeing Loop
The research record extends beyond Trials 001–003. The system has demonstrated persistent cross-run state and bounded cases in which retained outcomes changed later controller or selection behavior. It has not demonstrated autonomous model-weight learning or robust generalized performance improvement; the harder preregistered V3 replication failed its stated threshold.
Completed research
Status · SUPPORTED · METHOD UPDATE · July 23, 2026
Question: Can editorial selection begin from a current unresolved state, follow connected authoritative evidence, and produce a traceable recommendation without defaulting to a blank prompt or the easiest derivative?
Bounded conclusion: The bounded breadcrumb method works as implemented: it starts from a trigger, follows traceable evidence, constrains model interpretation, records alternatives, and preserves human authority. The proof does not establish that it produces better editorial choices than every competing retrieval method.
Read the completed research record →
Status · SUPPORTED · RESEARCH REPORT · July 23, 2026
Question: Does a prior result alter the next selection only when it carries specific, relevant pressure rather than vague or irrelevant history?
Bounded conclusion: Within the bounded harness, a specific relevant prior result changed future selection while vague feedback did not. This supports the operational selection-change criterion but does not demonstrate consciousness, sentience, or general autonomous learning.
Read the completed research record →
Status · SUPPORTED · RESEARCH REPORT · July 23, 2026
Question: Does bounded breadcrumb traversal plus contradiction search recover materially more relevant evidence than nearest-page retrieval on one long document?
Bounded conclusion: For this document and question, bounded breadcrumb traversal recovered more declared evidence than nearest-page retrieval, and contradiction search recovered one additional limit at greater context cost. The run supports deeper retrieval as a candidate method, not a universal optimal stopping rule.
Read the completed research record →
Status · SUPPORTED · REPLICATION · July 23, 2026
Question: Can Review Lab distinguish what an operational research report attempted, what worked, what was established, and what remains unresolved?
Bounded conclusion: In this run, Review Lab correctly preserved the distinction between a technical effect and an earned quality improvement, while retaining the report’s missing evidence and unresolved confounds. It validated claim calibration, not the underlying audio result.
Read the completed research record →
Status · SUPPORTED · METHOD UPDATE · July 23, 2026
Question: Can Review Lab prevent incomplete source retrieval from becoming an overconfident review conclusion?
Bounded conclusion: Review Lab can deterministically prevent incomplete source coverage from receiving its highest experimental review status, and its production prompt now prevents missing retrieved counterevidence from being described as nonexistent. This improves calibration but does not establish overall review superiority.
Read the completed research record →
Public thesis
- The human provides the question, purpose, imagination, dissatisfaction, and judgment.
- The language model helps expand, structure, test, and operationalize the idea.
- A fluent answer is not proof.
- A prototype is not a theory.
- A repeated claim is not independent evidence.
- A failed experiment is not wasted work.
- An unresolved question is a legitimate research result.
- Automation of execution is not automation of intention.
Return to Transit
